xref: /dflybsd-src/sys/kern/vfs_aio.c (revision b5b0912b1891e95ccc48cad83f09239ccb7ffc16)
1 /*
2  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. John S. Dyson's name may not be used to endorse or promote products
10  *    derived from this software without specific prior written permission.
11  *
12  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
13  * bad that happens because of using this software isn't the responsibility
14  * of the author.  This software is distributed AS-IS.
15  *
16  * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
17  * $DragonFly: src/sys/kern/vfs_aio.c,v 1.15 2005/06/06 15:02:28 dillon Exp $
18  */
19 
20 /*
21  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
22  */
23 
24 #include <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/buf.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
31 #include <sys/file.h>
32 #include <sys/lock.h>
33 #include <sys/unistd.h>
34 #include <sys/proc.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
41 #include <sys/conf.h>
42 #include <sys/event.h>
43 
44 #include <vm/vm.h>
45 #include <vm/vm_extern.h>
46 #include <vm/pmap.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
49 #include <sys/aio.h>
50 #include <sys/file2.h>
51 #include <sys/buf2.h>
52 
53 #include <machine/limits.h>
54 #include "opt_vfs_aio.h"
55 
56 #ifdef VFS_AIO
57 
58 /*
59  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
60  * overflow.
61  */
62 static	long jobrefid;
63 
64 #define JOBST_NULL		0x0
65 #define JOBST_JOBQGLOBAL	0x2
66 #define JOBST_JOBRUNNING	0x3
67 #define JOBST_JOBFINISHED	0x4
68 #define	JOBST_JOBQBUF		0x5
69 #define	JOBST_JOBBFINISHED	0x6
70 
71 #ifndef MAX_AIO_PER_PROC
72 #define MAX_AIO_PER_PROC	32
73 #endif
74 
75 #ifndef MAX_AIO_QUEUE_PER_PROC
76 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
77 #endif
78 
79 #ifndef MAX_AIO_PROCS
80 #define MAX_AIO_PROCS		32
81 #endif
82 
83 #ifndef MAX_AIO_QUEUE
84 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
85 #endif
86 
87 #ifndef TARGET_AIO_PROCS
88 #define TARGET_AIO_PROCS	4
89 #endif
90 
91 #ifndef MAX_BUF_AIO
92 #define MAX_BUF_AIO		16
93 #endif
94 
95 #ifndef AIOD_TIMEOUT_DEFAULT
96 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
97 #endif
98 
99 #ifndef AIOD_LIFETIME_DEFAULT
100 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
101 #endif
102 
103 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
104 
105 static int max_aio_procs = MAX_AIO_PROCS;
106 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
107 	CTLFLAG_RW, &max_aio_procs, 0,
108 	"Maximum number of kernel threads to use for handling async IO");
109 
110 static int num_aio_procs = 0;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
112 	CTLFLAG_RD, &num_aio_procs, 0,
113 	"Number of presently active kernel threads for async IO");
114 
115 /*
116  * The code will adjust the actual number of AIO processes towards this
117  * number when it gets a chance.
118  */
119 static int target_aio_procs = TARGET_AIO_PROCS;
120 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
121 	0, "Preferred number of ready kernel threads for async IO");
122 
123 static int max_queue_count = MAX_AIO_QUEUE;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
125     "Maximum number of aio requests to queue, globally");
126 
127 static int num_queue_count = 0;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
129     "Number of queued aio requests");
130 
131 static int num_buf_aio = 0;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
133     "Number of aio requests presently handled by the buf subsystem");
134 
135 /* Number of async I/O thread in the process of being started */
136 /* XXX This should be local to _aio_aqueue() */
137 static int num_aio_resv_start = 0;
138 
139 static int aiod_timeout;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
141     "Timeout value for synchronous aio operations");
142 
143 static int aiod_lifetime;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
145     "Maximum lifetime for idle aiod");
146 
147 static int max_aio_per_proc = MAX_AIO_PER_PROC;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
149     0, "Maximum active aio requests per process (stored in the process)");
150 
151 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
153     &max_aio_queue_per_proc, 0,
154     "Maximum queued aio requests per process (stored in the process)");
155 
156 static int max_buf_aio = MAX_BUF_AIO;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
158     "Maximum buf aio requests per process (stored in the process)");
159 
160 /*
161  * AIO process info
162  */
163 #define AIOP_FREE	0x1			/* proc on free queue */
164 #define AIOP_SCHED	0x2			/* proc explicitly scheduled */
165 
166 struct aioproclist {
167 	int aioprocflags;			/* AIO proc flags */
168 	TAILQ_ENTRY(aioproclist) list;		/* List of processes */
169 	struct proc *aioproc;			/* The AIO thread */
170 };
171 
172 /*
173  * data-structure for lio signal management
174  */
175 struct aio_liojob {
176 	int	lioj_flags;
177 	int	lioj_buffer_count;
178 	int	lioj_buffer_finished_count;
179 	int	lioj_queue_count;
180 	int	lioj_queue_finished_count;
181 	struct	sigevent lioj_signal;	/* signal on all I/O done */
182 	TAILQ_ENTRY(aio_liojob) lioj_list;
183 	struct	kaioinfo *lioj_ki;
184 };
185 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
186 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
187 
188 /*
189  * per process aio data structure
190  */
191 struct kaioinfo {
192 	int	kaio_flags;		/* per process kaio flags */
193 	int	kaio_maxactive_count;	/* maximum number of AIOs */
194 	int	kaio_active_count;	/* number of currently used AIOs */
195 	int	kaio_qallowed_count;	/* maxiumu size of AIO queue */
196 	int	kaio_queue_count;	/* size of AIO queue */
197 	int	kaio_ballowed_count;	/* maximum number of buffers */
198 	int	kaio_queue_finished_count; /* number of daemon jobs finished */
199 	int	kaio_buffer_count;	/* number of physio buffers */
200 	int	kaio_buffer_finished_count; /* count of I/O done */
201 	struct 	proc *kaio_p;		/* process that uses this kaio block */
202 	TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
203 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* job queue for process */
204 	TAILQ_HEAD(,aiocblist) kaio_jobdone;	/* done queue for process */
205 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* buffer job queue for process */
206 	TAILQ_HEAD(,aiocblist) kaio_bufdone;	/* buffer done queue for process */
207 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;	/* queue for aios waiting on sockets */
208 };
209 
210 #define KAIO_RUNDOWN	0x1	/* process is being run down */
211 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
212 
213 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
214 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* Async job list */
215 static TAILQ_HEAD(,aiocblist) aio_bufjobs;		/* Phys I/O job list */
216 static TAILQ_HEAD(,aiocblist) aio_freejobs;		/* Pool of free jobs */
217 
218 static void	aio_init_aioinfo(struct proc *p);
219 static void	aio_onceonly(void *);
220 static int	aio_free_entry(struct aiocblist *aiocbe);
221 static void	aio_process(struct aiocblist *aiocbe);
222 static int	aio_newproc(void);
223 static int	aio_aqueue(struct aiocb *job, int type);
224 static void	aio_physwakeup(struct buf *bp);
225 static int	aio_fphysio(struct aiocblist *aiocbe);
226 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
227 static void	aio_daemon(void *uproc);
228 static void	process_signal(void *aioj);
229 
230 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
231 
232 /*
233  * Zones for:
234  * 	kaio	Per process async io info
235  *	aiop	async io thread data
236  *	aiocb	async io jobs
237  *	aiol	list io job pointer - internal to aio_suspend XXX
238  *	aiolio	list io jobs
239  */
240 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
241 
242 /*
243  * Startup initialization
244  */
245 static void
246 aio_onceonly(void *na)
247 {
248 	TAILQ_INIT(&aio_freeproc);
249 	TAILQ_INIT(&aio_activeproc);
250 	TAILQ_INIT(&aio_jobs);
251 	TAILQ_INIT(&aio_bufjobs);
252 	TAILQ_INIT(&aio_freejobs);
253 	kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
254 	aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
255 	aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
256 	aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
257 	aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
258 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
259 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
260 	jobrefid = 1;
261 }
262 
263 /*
264  * Init the per-process aioinfo structure.  The aioinfo limits are set
265  * per-process for user limit (resource) management.
266  */
267 static void
268 aio_init_aioinfo(struct proc *p)
269 {
270 	struct kaioinfo *ki;
271 	if (p->p_aioinfo == NULL) {
272 		ki = zalloc(kaio_zone);
273 		p->p_aioinfo = ki;
274 		ki->kaio_flags = 0;
275 		ki->kaio_maxactive_count = max_aio_per_proc;
276 		ki->kaio_active_count = 0;
277 		ki->kaio_qallowed_count = max_aio_queue_per_proc;
278 		ki->kaio_queue_count = 0;
279 		ki->kaio_ballowed_count = max_buf_aio;
280 		ki->kaio_buffer_count = 0;
281 		ki->kaio_buffer_finished_count = 0;
282 		ki->kaio_p = p;
283 		TAILQ_INIT(&ki->kaio_jobdone);
284 		TAILQ_INIT(&ki->kaio_jobqueue);
285 		TAILQ_INIT(&ki->kaio_bufdone);
286 		TAILQ_INIT(&ki->kaio_bufqueue);
287 		TAILQ_INIT(&ki->kaio_liojoblist);
288 		TAILQ_INIT(&ki->kaio_sockqueue);
289 	}
290 
291 	while (num_aio_procs < target_aio_procs)
292 		aio_newproc();
293 }
294 
295 /*
296  * Free a job entry.  Wait for completion if it is currently active, but don't
297  * delay forever.  If we delay, we return a flag that says that we have to
298  * restart the queue scan.
299  */
300 static int
301 aio_free_entry(struct aiocblist *aiocbe)
302 {
303 	struct kaioinfo *ki;
304 	struct aio_liojob *lj;
305 	struct proc *p;
306 	int error;
307 
308 	if (aiocbe->jobstate == JOBST_NULL)
309 		panic("aio_free_entry: freeing already free job");
310 
311 	p = aiocbe->userproc;
312 	ki = p->p_aioinfo;
313 	lj = aiocbe->lio;
314 	if (ki == NULL)
315 		panic("aio_free_entry: missing p->p_aioinfo");
316 
317 	while (aiocbe->jobstate == JOBST_JOBRUNNING) {
318 		aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
319 		tsleep(aiocbe, 0, "jobwai", 0);
320 	}
321 	if (aiocbe->bp == NULL) {
322 		if (ki->kaio_queue_count <= 0)
323 			panic("aio_free_entry: process queue size <= 0");
324 		if (num_queue_count <= 0)
325 			panic("aio_free_entry: system wide queue size <= 0");
326 
327 		if (lj) {
328 			lj->lioj_queue_count--;
329 			if (aiocbe->jobflags & AIOCBLIST_DONE)
330 				lj->lioj_queue_finished_count--;
331 		}
332 		ki->kaio_queue_count--;
333 		if (aiocbe->jobflags & AIOCBLIST_DONE)
334 			ki->kaio_queue_finished_count--;
335 		num_queue_count--;
336 	} else {
337 		if (lj) {
338 			lj->lioj_buffer_count--;
339 			if (aiocbe->jobflags & AIOCBLIST_DONE)
340 				lj->lioj_buffer_finished_count--;
341 		}
342 		if (aiocbe->jobflags & AIOCBLIST_DONE)
343 			ki->kaio_buffer_finished_count--;
344 		ki->kaio_buffer_count--;
345 		num_buf_aio--;
346 	}
347 
348 	/* aiocbe is going away, we need to destroy any knotes */
349 	knote_remove(p->p_thread, &aiocbe->klist);
350 
351 	if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
352 	    && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
353 		ki->kaio_flags &= ~KAIO_WAKEUP;
354 		wakeup(p);
355 	}
356 
357 	if (aiocbe->jobstate == JOBST_JOBQBUF) {
358 		if ((error = aio_fphysio(aiocbe)) != 0)
359 			return error;
360 		if (aiocbe->jobstate != JOBST_JOBBFINISHED)
361 			panic("aio_free_entry: invalid physio finish-up state");
362 		crit_enter();
363 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
364 		crit_exit();
365 	} else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
366 		crit_enter();
367 		TAILQ_REMOVE(&aio_jobs, aiocbe, list);
368 		TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
369 		crit_exit();
370 	} else if (aiocbe->jobstate == JOBST_JOBFINISHED)
371 		TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
372 	else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
373 		crit_enter();
374 		TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
375 		crit_exit();
376 		if (aiocbe->bp) {
377 			vunmapbuf(aiocbe->bp);
378 			relpbuf(aiocbe->bp, NULL);
379 			aiocbe->bp = NULL;
380 		}
381 	}
382 	if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
383 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
384 		zfree(aiolio_zone, lj);
385 	}
386 	aiocbe->jobstate = JOBST_NULL;
387 	callout_stop(&aiocbe->timeout);
388 	fdrop(aiocbe->fd_file, curthread);
389 	TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
390 	return 0;
391 }
392 #endif /* VFS_AIO */
393 
394 /*
395  * Rundown the jobs for a given process.
396  */
397 void
398 aio_proc_rundown(struct proc *p)
399 {
400 #ifndef VFS_AIO
401 	return;
402 #else
403 	struct kaioinfo *ki;
404 	struct aio_liojob *lj, *ljn;
405 	struct aiocblist *aiocbe, *aiocbn;
406 	struct file *fp;
407 	struct socket *so;
408 
409 	ki = p->p_aioinfo;
410 	if (ki == NULL)
411 		return;
412 
413 	ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
414 	while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
415 	    ki->kaio_buffer_finished_count)) {
416 		ki->kaio_flags |= KAIO_RUNDOWN;
417 		if (tsleep(p, 0, "kaiowt", aiod_timeout))
418 			break;
419 	}
420 
421 	/*
422 	 * Move any aio ops that are waiting on socket I/O to the normal job
423 	 * queues so they are cleaned up with any others.
424 	 */
425 	crit_enter();
426 	for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
427 	    aiocbn) {
428 		aiocbn = TAILQ_NEXT(aiocbe, plist);
429 		fp = aiocbe->fd_file;
430 		if (fp != NULL) {
431 			so = (struct socket *)fp->f_data;
432 			TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
433 			if (TAILQ_EMPTY(&so->so_aiojobq)) {
434 				so->so_snd.sb_flags &= ~SB_AIO;
435 				so->so_rcv.sb_flags &= ~SB_AIO;
436 			}
437 		}
438 		TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
439 		TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
440 		TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
441 	}
442 	crit_exit();
443 
444 restart1:
445 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
446 		aiocbn = TAILQ_NEXT(aiocbe, plist);
447 		if (aio_free_entry(aiocbe))
448 			goto restart1;
449 	}
450 
451 restart2:
452 	for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
453 	    aiocbn) {
454 		aiocbn = TAILQ_NEXT(aiocbe, plist);
455 		if (aio_free_entry(aiocbe))
456 			goto restart2;
457 	}
458 
459 restart3:
460 	crit_enter();
461 	while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
462 		ki->kaio_flags |= KAIO_WAKEUP;
463 		tsleep(p, 0, "aioprn", 0);
464 		crit_exit();
465 		goto restart3;
466 	}
467 	crit_exit();
468 
469 restart4:
470 	crit_enter();
471 	for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
472 		aiocbn = TAILQ_NEXT(aiocbe, plist);
473 		if (aio_free_entry(aiocbe)) {
474 			crit_exit();
475 			goto restart4;
476 		}
477 	}
478 	crit_exit();
479 
480         /*
481          * If we've slept, jobs might have moved from one queue to another.
482          * Retry rundown if we didn't manage to empty the queues.
483          */
484         if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
485 	    TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
486 	    TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
487 	    TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
488 		goto restart1;
489 
490 	for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
491 		ljn = TAILQ_NEXT(lj, lioj_list);
492 		if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
493 		    0)) {
494 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
495 			zfree(aiolio_zone, lj);
496 		} else {
497 #ifdef DIAGNOSTIC
498 			printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
499 			    "QF:%d\n", lj->lioj_buffer_count,
500 			    lj->lioj_buffer_finished_count,
501 			    lj->lioj_queue_count,
502 			    lj->lioj_queue_finished_count);
503 #endif
504 		}
505 	}
506 
507 	zfree(kaio_zone, ki);
508 	p->p_aioinfo = NULL;
509 #endif /* VFS_AIO */
510 }
511 
512 #ifdef VFS_AIO
513 /*
514  * Select a job to run (called by an AIO daemon).
515  */
516 static struct aiocblist *
517 aio_selectjob(struct aioproclist *aiop)
518 {
519 	struct aiocblist *aiocbe;
520 	struct kaioinfo *ki;
521 	struct proc *userp;
522 
523 	crit_enter();
524 	for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
525 	    TAILQ_NEXT(aiocbe, list)) {
526 		userp = aiocbe->userproc;
527 		ki = userp->p_aioinfo;
528 
529 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
530 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
531 			crit_exit();
532 			return aiocbe;
533 		}
534 	}
535 	crit_exit();
536 
537 	return NULL;
538 }
539 
540 /*
541  * The AIO processing activity.  This is the code that does the I/O request for
542  * the non-physio version of the operations.  The normal vn operations are used,
543  * and this code should work in all instances for every type of file, including
544  * pipes, sockets, fifos, and regular files.
545  */
546 static void
547 aio_process(struct aiocblist *aiocbe)
548 {
549 	struct thread *mytd;
550 	struct aiocb *cb;
551 	struct file *fp;
552 	struct uio auio;
553 	struct iovec aiov;
554 	int cnt;
555 	int error;
556 	int oublock_st, oublock_end;
557 	int inblock_st, inblock_end;
558 
559 	mytd = curthread;
560 	cb = &aiocbe->uaiocb;
561 	fp = aiocbe->fd_file;
562 
563 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
564 	aiov.iov_len = cb->aio_nbytes;
565 
566 	auio.uio_iov = &aiov;
567 	auio.uio_iovcnt = 1;
568 	auio.uio_offset = cb->aio_offset;
569 	auio.uio_resid = cb->aio_nbytes;
570 	cnt = cb->aio_nbytes;
571 	auio.uio_segflg = UIO_USERSPACE;
572 	auio.uio_td = mytd;
573 
574 	inblock_st = mytd->td_proc->p_stats->p_ru.ru_inblock;
575 	oublock_st = mytd->td_proc->p_stats->p_ru.ru_oublock;
576 	/*
577 	 * _aio_aqueue() acquires a reference to the file that is
578 	 * released in aio_free_entry().
579 	 */
580 	if (cb->aio_lio_opcode == LIO_READ) {
581 		auio.uio_rw = UIO_READ;
582 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
583 	} else {
584 		auio.uio_rw = UIO_WRITE;
585 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
586 	}
587 	inblock_end = mytd->td_proc->p_stats->p_ru.ru_inblock;
588 	oublock_end = mytd->td_proc->p_stats->p_ru.ru_oublock;
589 
590 	aiocbe->inputcharge = inblock_end - inblock_st;
591 	aiocbe->outputcharge = oublock_end - oublock_st;
592 
593 	if ((error) && (auio.uio_resid != cnt)) {
594 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
595 			error = 0;
596 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
597 			psignal(aiocbe->userproc, SIGPIPE);
598 	}
599 
600 	cnt -= auio.uio_resid;
601 	cb->_aiocb_private.error = error;
602 	cb->_aiocb_private.status = cnt;
603 }
604 
605 /*
606  * The AIO daemon, most of the actual work is done in aio_process,
607  * but the setup (and address space mgmt) is done in this routine.
608  *
609  * The MP lock is held on entry.
610  */
611 static void
612 aio_daemon(void *uproc)
613 {
614 	struct aio_liojob *lj;
615 	struct aiocb *cb;
616 	struct aiocblist *aiocbe;
617 	struct aioproclist *aiop;
618 	struct kaioinfo *ki;
619 	struct proc *curcp, *mycp, *userp;
620 	struct vmspace *myvm, *tmpvm;
621 	struct ucred *cr;
622 
623 	/*
624 	 * Local copies of curproc (cp) and vmspace (myvm)
625 	 */
626 	mycp = curproc;
627 	myvm = mycp->p_vmspace;
628 
629 	if (mycp->p_textvp) {
630 		vrele(mycp->p_textvp);
631 		mycp->p_textvp = NULL;
632 	}
633 
634 	/*
635 	 * Allocate and ready the aio control info.  There is one aiop structure
636 	 * per daemon.
637 	 */
638 	aiop = zalloc(aiop_zone);
639 	aiop->aioproc = mycp;
640 	aiop->aioprocflags |= AIOP_FREE;
641 
642 	crit_enter();
643 
644 	/*
645 	 * Place thread (lightweight process) onto the AIO free thread list.
646 	 */
647 	if (TAILQ_EMPTY(&aio_freeproc))
648 		wakeup(&aio_freeproc);
649 	TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
650 
651 	crit_exit();
652 
653 	/* Make up a name for the daemon. */
654 	strcpy(mycp->p_comm, "aiod");
655 
656 	/*
657 	 * Get rid of our current filedescriptors.  AIOD's don't need any
658 	 * filedescriptors, except as temporarily inherited from the client.
659 	 * Credentials are also cloned, and made equivalent to "root".
660 	 */
661 	fdfree(mycp);
662 	mycp->p_fd = NULL;
663 	cr = cratom(&mycp->p_ucred);
664 	cr->cr_uid = 0;
665 	uireplace(&cr->cr_uidinfo, uifind(0));
666 	cr->cr_ngroups = 1;
667 	cr->cr_groups[0] = 1;
668 
669 	/* The daemon resides in its own pgrp. */
670 	enterpgrp(mycp, mycp->p_pid, 1);
671 
672 	/* Mark special process type. */
673 	mycp->p_flag |= P_SYSTEM | P_KTHREADP;
674 
675 	/*
676 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
677 	 * and creating too many daemons.)
678 	 */
679 	wakeup(mycp);
680 
681 	for (;;) {
682 		/*
683 		 * curcp is the current daemon process context.
684 		 * userp is the current user process context.
685 		 */
686 		curcp = mycp;
687 
688 		/*
689 		 * Take daemon off of free queue
690 		 */
691 		if (aiop->aioprocflags & AIOP_FREE) {
692 			crit_enter();
693 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
694 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
695 			aiop->aioprocflags &= ~AIOP_FREE;
696 			crit_exit();
697 		}
698 		aiop->aioprocflags &= ~AIOP_SCHED;
699 
700 		/*
701 		 * Check for jobs.
702 		 */
703 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
704 			cb = &aiocbe->uaiocb;
705 			userp = aiocbe->userproc;
706 
707 			aiocbe->jobstate = JOBST_JOBRUNNING;
708 
709 			/*
710 			 * Connect to process address space for user program.
711 			 */
712 			if (userp != curcp) {
713 				/*
714 				 * Save the current address space that we are
715 				 * connected to.
716 				 */
717 				tmpvm = mycp->p_vmspace;
718 
719 				/*
720 				 * Point to the new user address space, and
721 				 * refer to it.
722 				 */
723 				mycp->p_vmspace = userp->p_vmspace;
724 				mycp->p_vmspace->vm_refcnt++;
725 
726 				/* Activate the new mapping. */
727 				pmap_activate(mycp);
728 
729 				/*
730 				 * If the old address space wasn't the daemons
731 				 * own address space, then we need to remove the
732 				 * daemon's reference from the other process
733 				 * that it was acting on behalf of.
734 				 */
735 				if (tmpvm != myvm) {
736 					vmspace_free(tmpvm);
737 				}
738 				curcp = userp;
739 			}
740 
741 			ki = userp->p_aioinfo;
742 			lj = aiocbe->lio;
743 
744 			/* Account for currently active jobs. */
745 			ki->kaio_active_count++;
746 
747 			/* Do the I/O function. */
748 			aio_process(aiocbe);
749 
750 			/* Decrement the active job count. */
751 			ki->kaio_active_count--;
752 
753 			/*
754 			 * Increment the completion count for wakeup/signal
755 			 * comparisons.
756 			 */
757 			aiocbe->jobflags |= AIOCBLIST_DONE;
758 			ki->kaio_queue_finished_count++;
759 			if (lj)
760 				lj->lioj_queue_finished_count++;
761 			if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
762 			    & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
763 				ki->kaio_flags &= ~KAIO_WAKEUP;
764 				wakeup(userp);
765 			}
766 
767 			crit_enter();
768 			if (lj && (lj->lioj_flags &
769 			    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
770 				if ((lj->lioj_queue_finished_count ==
771 				    lj->lioj_queue_count) &&
772 				    (lj->lioj_buffer_finished_count ==
773 				    lj->lioj_buffer_count)) {
774 						psignal(userp,
775 						    lj->lioj_signal.sigev_signo);
776 						lj->lioj_flags |=
777 						    LIOJ_SIGNAL_POSTED;
778 				}
779 			}
780 			crit_exit();
781 
782 			aiocbe->jobstate = JOBST_JOBFINISHED;
783 
784 			crit_enter();
785 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
786 			TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
787 			crit_exit();
788 			KNOTE(&aiocbe->klist, 0);
789 
790 			if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
791 				wakeup(aiocbe);
792 				aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
793 			}
794 
795 			if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
796 				psignal(userp, cb->aio_sigevent.sigev_signo);
797 			}
798 		}
799 
800 		/*
801 		 * Disconnect from user address space.
802 		 */
803 		if (curcp != mycp) {
804 			/* Get the user address space to disconnect from. */
805 			tmpvm = mycp->p_vmspace;
806 
807 			/* Get original address space for daemon. */
808 			mycp->p_vmspace = myvm;
809 
810 			/* Activate the daemon's address space. */
811 			pmap_activate(mycp);
812 #ifdef DIAGNOSTIC
813 			if (tmpvm == myvm) {
814 				printf("AIOD: vmspace problem -- %d\n",
815 				    mycp->p_pid);
816 			}
817 #endif
818 			/* Remove our vmspace reference. */
819 			vmspace_free(tmpvm);
820 
821 			curcp = mycp;
822 		}
823 
824 		/*
825 		 * If we are the first to be put onto the free queue, wakeup
826 		 * anyone waiting for a daemon.
827 		 */
828 		crit_enter();
829 		TAILQ_REMOVE(&aio_activeproc, aiop, list);
830 		if (TAILQ_EMPTY(&aio_freeproc))
831 			wakeup(&aio_freeproc);
832 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
833 		aiop->aioprocflags |= AIOP_FREE;
834 		crit_exit();
835 
836 		/*
837 		 * If daemon is inactive for a long time, allow it to exit,
838 		 * thereby freeing resources.
839 		 */
840 		if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
841 		    0, "aiordy", aiod_lifetime)) {
842 			crit_enter();
843 			if (TAILQ_EMPTY(&aio_jobs)) {
844 				if ((aiop->aioprocflags & AIOP_FREE) &&
845 				    (num_aio_procs > target_aio_procs)) {
846 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
847 					crit_exit();
848 					zfree(aiop_zone, aiop);
849 					num_aio_procs--;
850 #ifdef DIAGNOSTIC
851 					if (mycp->p_vmspace->vm_refcnt <= 1) {
852 						printf("AIOD: bad vm refcnt for"
853 						    " exiting daemon: %d\n",
854 						    mycp->p_vmspace->vm_refcnt);
855 					}
856 #endif
857 					exit1(0);
858 				}
859 			}
860 			crit_exit();
861 		}
862 	}
863 }
864 
865 /*
866  * Create a new AIO daemon.  This is mostly a kernel-thread fork routine.  The
867  * AIO daemon modifies its environment itself.
868  */
869 static int
870 aio_newproc()
871 {
872 	int error;
873 	struct proc *p, *np;
874 
875 	p = &proc0;
876 	error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
877 	if (error)
878 		return error;
879 	cpu_set_fork_handler(np, aio_daemon, curproc);
880 	start_forked_proc(p, np);
881 
882 	/*
883 	 * Wait until daemon is started, but continue on just in case to
884 	 * handle error conditions.
885 	 */
886 	error = tsleep(np, 0, "aiosta", aiod_timeout);
887 	num_aio_procs++;
888 
889 	return error;
890 }
891 
892 /*
893  * Try the high-performance, low-overhead physio method for eligible
894  * VCHR devices.  This method doesn't use an aio helper thread, and
895  * thus has very low overhead.
896  *
897  * Assumes that the caller, _aio_aqueue(), has incremented the file
898  * structure's reference count, preventing its deallocation for the
899  * duration of this call.
900  */
901 static int
902 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
903 {
904 	int error;
905 	struct aiocb *cb;
906 	struct file *fp;
907 	struct buf *bp;
908 	struct vnode *vp;
909 	struct kaioinfo *ki;
910 	struct aio_liojob *lj;
911 	int notify;
912 
913 	cb = &aiocbe->uaiocb;
914 	fp = aiocbe->fd_file;
915 
916 	if (fp->f_type != DTYPE_VNODE)
917 		return (-1);
918 
919 	vp = (struct vnode *)fp->f_data;
920 
921 	/*
922 	 * If its not a disk, we don't want to return a positive error.
923 	 * It causes the aio code to not fall through to try the thread
924 	 * way when you're talking to a regular file.
925 	 */
926 	if (!vn_isdisk(vp, &error)) {
927 		if (error == ENOTBLK)
928 			return (-1);
929 		else
930 			return (error);
931 	}
932 
933  	if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
934 		return (-1);
935 
936 	if (cb->aio_nbytes >
937 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
938 		return (-1);
939 
940 	ki = p->p_aioinfo;
941 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
942 		return (-1);
943 
944 	ki->kaio_buffer_count++;
945 
946 	lj = aiocbe->lio;
947 	if (lj)
948 		lj->lioj_buffer_count++;
949 
950 	/* Create and build a buffer header for a transfer. */
951 	bp = (struct buf *)getpbuf(NULL);
952 	BUF_KERNPROC(bp);
953 
954 	/*
955 	 * Get a copy of the kva from the physical buffer.
956 	 */
957 	bp->b_caller1 = p;
958 	bp->b_dev = vp->v_rdev;
959 	error = 0;
960 
961 	bp->b_bcount = cb->aio_nbytes;
962 	bp->b_bufsize = cb->aio_nbytes;
963 	bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
964 	    B_WRITE : B_READ);
965 	bp->b_iodone = aio_physwakeup;
966 	bp->b_saveaddr = bp->b_data;
967 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
968 	bp->b_blkno = btodb(cb->aio_offset);
969 
970 	/* Bring buffer into kernel space. */
971 	if (vmapbuf(bp) < 0) {
972 		error = EFAULT;
973 		goto doerror;
974 	}
975 
976 	crit_enter();
977 
978 	aiocbe->bp = bp;
979 	bp->b_spc = (void *)aiocbe;
980 	TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
981 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
982 	aiocbe->jobstate = JOBST_JOBQBUF;
983 	cb->_aiocb_private.status = cb->aio_nbytes;
984 	num_buf_aio++;
985 	bp->b_error = 0;
986 
987 	crit_exit();
988 
989 	/* Perform transfer. */
990 	BUF_STRATEGY(bp, 0);
991 
992 	notify = 0;
993 	crit_enter();
994 
995 	/*
996 	 * If we had an error invoking the request, or an error in processing
997 	 * the request before we have returned, we process it as an error in
998 	 * transfer.  Note that such an I/O error is not indicated immediately,
999 	 * but is returned using the aio_error mechanism.  In this case,
1000 	 * aio_suspend will return immediately.
1001 	 */
1002 	if (bp->b_error || (bp->b_flags & B_ERROR)) {
1003 		struct aiocb *job = aiocbe->uuaiocb;
1004 
1005 		aiocbe->uaiocb._aiocb_private.status = 0;
1006 		suword(&job->_aiocb_private.status, 0);
1007 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1008 		suword(&job->_aiocb_private.error, bp->b_error);
1009 
1010 		ki->kaio_buffer_finished_count++;
1011 
1012 		if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1013 			aiocbe->jobstate = JOBST_JOBBFINISHED;
1014 			aiocbe->jobflags |= AIOCBLIST_DONE;
1015 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1016 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1017 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1018 			notify = 1;
1019 		}
1020 	}
1021 	crit_exit();
1022 	if (notify)
1023 		KNOTE(&aiocbe->klist, 0);
1024 	return 0;
1025 
1026 doerror:
1027 	ki->kaio_buffer_count--;
1028 	if (lj)
1029 		lj->lioj_buffer_count--;
1030 	aiocbe->bp = NULL;
1031 	relpbuf(bp, NULL);
1032 	return error;
1033 }
1034 
1035 /*
1036  * This waits/tests physio completion.
1037  */
1038 static int
1039 aio_fphysio(struct aiocblist *iocb)
1040 {
1041 	struct buf *bp;
1042 	int error;
1043 
1044 	bp = iocb->bp;
1045 
1046 	crit_enter();
1047 	while ((bp->b_flags & B_DONE) == 0) {
1048 		if (tsleep(bp, 0, "physstr", aiod_timeout)) {
1049 			if ((bp->b_flags & B_DONE) == 0) {
1050 				crit_exit();
1051 				return EINPROGRESS;
1052 			} else
1053 				break;
1054 		}
1055 	}
1056 	crit_exit();
1057 
1058 	/* Release mapping into kernel space. */
1059 	vunmapbuf(bp);
1060 	iocb->bp = 0;
1061 
1062 	error = 0;
1063 
1064 	/* Check for an error. */
1065 	if (bp->b_flags & B_ERROR)
1066 		error = bp->b_error;
1067 
1068 	relpbuf(bp, NULL);
1069 	return (error);
1070 }
1071 #endif /* VFS_AIO */
1072 
1073 /*
1074  * Wake up aio requests that may be serviceable now.
1075  */
1076 void
1077 aio_swake(struct socket *so, struct sockbuf *sb)
1078 {
1079 #ifndef VFS_AIO
1080 	return;
1081 #else
1082 	struct aiocblist *cb,*cbn;
1083 	struct proc *p;
1084 	struct kaioinfo *ki = NULL;
1085 	int opcode, wakecount = 0;
1086 	struct aioproclist *aiop;
1087 
1088 	if (sb == &so->so_snd) {
1089 		opcode = LIO_WRITE;
1090 		so->so_snd.sb_flags &= ~SB_AIO;
1091 	} else {
1092 		opcode = LIO_READ;
1093 		so->so_rcv.sb_flags &= ~SB_AIO;
1094 	}
1095 
1096 	for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1097 		cbn = TAILQ_NEXT(cb, list);
1098 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1099 			p = cb->userproc;
1100 			ki = p->p_aioinfo;
1101 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1102 			TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1103 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1104 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1105 			wakecount++;
1106 			if (cb->jobstate != JOBST_JOBQGLOBAL)
1107 				panic("invalid queue value");
1108 		}
1109 	}
1110 
1111 	while (wakecount--) {
1112 		if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1113 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1114 			TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1115 			aiop->aioprocflags &= ~AIOP_FREE;
1116 			wakeup(aiop->aioproc);
1117 		}
1118 	}
1119 #endif /* VFS_AIO */
1120 }
1121 
1122 #ifdef VFS_AIO
1123 /*
1124  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1125  * technique is done in this code.
1126  */
1127 static int
1128 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1129 {
1130 	struct proc *p = curproc;
1131 	struct filedesc *fdp;
1132 	struct file *fp;
1133 	unsigned int fd;
1134 	struct socket *so;
1135 	int error;
1136 	int opcode, user_opcode;
1137 	struct aiocblist *aiocbe;
1138 	struct aioproclist *aiop;
1139 	struct kaioinfo *ki;
1140 	struct kevent kev;
1141 	struct kqueue *kq;
1142 	struct file *kq_fp;
1143 
1144 	if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1145 		TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1146 	else
1147 		aiocbe = zalloc (aiocb_zone);
1148 
1149 	aiocbe->inputcharge = 0;
1150 	aiocbe->outputcharge = 0;
1151 	callout_init(&aiocbe->timeout);
1152 	SLIST_INIT(&aiocbe->klist);
1153 
1154 	suword(&job->_aiocb_private.status, -1);
1155 	suword(&job->_aiocb_private.error, 0);
1156 	suword(&job->_aiocb_private.kernelinfo, -1);
1157 
1158 	error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1159 	if (error) {
1160 		suword(&job->_aiocb_private.error, error);
1161 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1162 		return error;
1163 	}
1164 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1165 	    !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1166 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1167 		return EINVAL;
1168 	}
1169 
1170 	/* Save userspace address of the job info. */
1171 	aiocbe->uuaiocb = job;
1172 
1173 	/* Get the opcode. */
1174 	user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1175 	if (type != LIO_NOP)
1176 		aiocbe->uaiocb.aio_lio_opcode = type;
1177 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1178 
1179 	/* Get the fd info for process. */
1180 	fdp = p->p_fd;
1181 
1182 	/*
1183 	 * Range check file descriptor.
1184 	 */
1185 	fd = aiocbe->uaiocb.aio_fildes;
1186 	if (fd >= fdp->fd_nfiles) {
1187 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1188 		if (type == 0)
1189 			suword(&job->_aiocb_private.error, EBADF);
1190 		return EBADF;
1191 	}
1192 
1193 	fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1194 	if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1195 	    0))) {
1196 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1197 		if (type == 0)
1198 			suword(&job->_aiocb_private.error, EBADF);
1199 		return EBADF;
1200 	}
1201 	fhold(fp);
1202 
1203 	if (aiocbe->uaiocb.aio_offset == -1LL) {
1204 		error = EINVAL;
1205 		goto aqueue_fail;
1206 	}
1207 	error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1208 	if (error) {
1209 		error = EINVAL;
1210 		goto aqueue_fail;
1211 	}
1212 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1213 	if (jobrefid == LONG_MAX)
1214 		jobrefid = 1;
1215 	else
1216 		jobrefid++;
1217 
1218 	if (opcode == LIO_NOP) {
1219 		fdrop(fp, p->p_thread);
1220 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1221 		if (type == 0) {
1222 			suword(&job->_aiocb_private.error, 0);
1223 			suword(&job->_aiocb_private.status, 0);
1224 			suword(&job->_aiocb_private.kernelinfo, 0);
1225 		}
1226 		return 0;
1227 	}
1228 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1229 		if (type == 0)
1230 			suword(&job->_aiocb_private.status, 0);
1231 		error = EINVAL;
1232 		goto aqueue_fail;
1233 	}
1234 
1235 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1236 		kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1237 		kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1238 	}
1239 	else {
1240 		/*
1241 		 * This method for requesting kevent-based notification won't
1242 		 * work on the alpha, since we're passing in a pointer
1243 		 * via aio_lio_opcode, which is an int.  Use the SIGEV_KEVENT-
1244 		 * based method instead.
1245 		 */
1246 		if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1247 		    user_opcode == LIO_WRITE)
1248 			goto no_kqueue;
1249 
1250 		error = copyin((struct kevent *)(uintptr_t)user_opcode,
1251 		    &kev, sizeof(kev));
1252 		if (error)
1253 			goto aqueue_fail;
1254 	}
1255 	if ((u_int)kev.ident >= fdp->fd_nfiles ||
1256 	    (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1257 	    (kq_fp->f_type != DTYPE_KQUEUE)) {
1258 		error = EBADF;
1259 		goto aqueue_fail;
1260 	}
1261 	kq = (struct kqueue *)kq_fp->f_data;
1262 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1263 	kev.filter = EVFILT_AIO;
1264 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1265 	kev.data = (intptr_t)aiocbe;
1266 	error = kqueue_register(kq, &kev, p->p_thread);
1267 aqueue_fail:
1268 	if (error) {
1269 		fdrop(fp, p->p_thread);
1270 		TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1271 		if (type == 0)
1272 			suword(&job->_aiocb_private.error, error);
1273 		goto done;
1274 	}
1275 no_kqueue:
1276 
1277 	suword(&job->_aiocb_private.error, EINPROGRESS);
1278 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1279 	aiocbe->userproc = p;
1280 	aiocbe->jobflags = 0;
1281 	aiocbe->lio = lj;
1282 	ki = p->p_aioinfo;
1283 
1284 	if (fp->f_type == DTYPE_SOCKET) {
1285 		/*
1286 		 * Alternate queueing for socket ops: Reach down into the
1287 		 * descriptor to get the socket data.  Then check to see if the
1288 		 * socket is ready to be read or written (based on the requested
1289 		 * operation).
1290 		 *
1291 		 * If it is not ready for io, then queue the aiocbe on the
1292 		 * socket, and set the flags so we get a call when sbnotify()
1293 		 * happens.
1294 		 */
1295 		so = (struct socket *)fp->f_data;
1296 		crit_enter();
1297 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1298 		    LIO_WRITE) && (!sowriteable(so)))) {
1299 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1300 			TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1301 			if (opcode == LIO_READ)
1302 				so->so_rcv.sb_flags |= SB_AIO;
1303 			else
1304 				so->so_snd.sb_flags |= SB_AIO;
1305 			aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1306 			ki->kaio_queue_count++;
1307 			num_queue_count++;
1308 			crit_exit();
1309 			error = 0;
1310 			goto done;
1311 		}
1312 		crit_exit();
1313 	}
1314 
1315 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1316 		goto done;
1317 	if (error > 0) {
1318 		suword(&job->_aiocb_private.status, 0);
1319 		aiocbe->uaiocb._aiocb_private.error = error;
1320 		suword(&job->_aiocb_private.error, error);
1321 		goto done;
1322 	}
1323 
1324 	/* No buffer for daemon I/O. */
1325 	aiocbe->bp = NULL;
1326 
1327 	ki->kaio_queue_count++;
1328 	if (lj)
1329 		lj->lioj_queue_count++;
1330 	crit_enter();
1331 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1332 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1333 	crit_exit();
1334 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1335 
1336 	num_queue_count++;
1337 	error = 0;
1338 
1339 	/*
1340 	 * If we don't have a free AIO process, and we are below our quota, then
1341 	 * start one.  Otherwise, depend on the subsequent I/O completions to
1342 	 * pick-up this job.  If we don't successfully create the new process
1343 	 * (thread) due to resource issues, we return an error for now (EAGAIN),
1344 	 * which is likely not the correct thing to do.
1345 	 */
1346 	crit_enter();
1347 retryproc:
1348 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1349 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1350 		TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1351 		aiop->aioprocflags &= ~AIOP_FREE;
1352 		wakeup(aiop->aioproc);
1353 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1354 	    ((ki->kaio_active_count + num_aio_resv_start) <
1355 	    ki->kaio_maxactive_count)) {
1356 		num_aio_resv_start++;
1357 		if ((error = aio_newproc()) == 0) {
1358 			num_aio_resv_start--;
1359 			goto retryproc;
1360 		}
1361 		num_aio_resv_start--;
1362 	}
1363 	crit_exit();
1364 done:
1365 	return error;
1366 }
1367 
1368 /*
1369  * This routine queues an AIO request, checking for quotas.
1370  */
1371 static int
1372 aio_aqueue(struct aiocb *job, int type)
1373 {
1374 	struct proc *p = curproc;
1375 	struct kaioinfo *ki;
1376 
1377 	if (p->p_aioinfo == NULL)
1378 		aio_init_aioinfo(p);
1379 
1380 	if (num_queue_count >= max_queue_count)
1381 		return EAGAIN;
1382 
1383 	ki = p->p_aioinfo;
1384 	if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1385 		return EAGAIN;
1386 
1387 	return _aio_aqueue(job, NULL, type);
1388 }
1389 #endif /* VFS_AIO */
1390 
1391 /*
1392  * Support the aio_return system call, as a side-effect, kernel resources are
1393  * released.
1394  */
1395 int
1396 aio_return(struct aio_return_args *uap)
1397 {
1398 #ifndef VFS_AIO
1399 	return ENOSYS;
1400 #else
1401 	struct proc *p = curproc;
1402 	long jobref;
1403 	struct aiocblist *cb, *ncb;
1404 	struct aiocb *ujob;
1405 	struct kaioinfo *ki;
1406 
1407 	ki = p->p_aioinfo;
1408 	if (ki == NULL)
1409 		return EINVAL;
1410 
1411 	ujob = uap->aiocbp;
1412 
1413 	jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1414 	if (jobref == -1 || jobref == 0)
1415 		return EINVAL;
1416 
1417 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1418 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1419 		    jobref) {
1420 			if (ujob == cb->uuaiocb) {
1421 				uap->sysmsg_result =
1422 				    cb->uaiocb._aiocb_private.status;
1423 			} else
1424 				uap->sysmsg_result = EFAULT;
1425 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1426 				p->p_stats->p_ru.ru_oublock +=
1427 				    cb->outputcharge;
1428 				cb->outputcharge = 0;
1429 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1430 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1431 				cb->inputcharge = 0;
1432 			}
1433 			aio_free_entry(cb);
1434 			return 0;
1435 		}
1436 	}
1437 	crit_enter();
1438 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1439 		ncb = TAILQ_NEXT(cb, plist);
1440 		if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1441 		    == jobref) {
1442 			crit_exit();
1443 			if (ujob == cb->uuaiocb) {
1444 				uap->sysmsg_result =
1445 				    cb->uaiocb._aiocb_private.status;
1446 			} else
1447 				uap->sysmsg_result = EFAULT;
1448 			aio_free_entry(cb);
1449 			return 0;
1450 		}
1451 	}
1452 	crit_exit();
1453 
1454 	return (EINVAL);
1455 #endif /* VFS_AIO */
1456 }
1457 
1458 /*
1459  * Allow a process to wakeup when any of the I/O requests are completed.
1460  */
1461 int
1462 aio_suspend(struct aio_suspend_args *uap)
1463 {
1464 #ifndef VFS_AIO
1465 	return ENOSYS;
1466 #else
1467 	struct proc *p = curproc;
1468 	struct timeval atv;
1469 	struct timespec ts;
1470 	struct aiocb *const *cbptr, *cbp;
1471 	struct kaioinfo *ki;
1472 	struct aiocblist *cb;
1473 	int i;
1474 	int njoblist;
1475 	int error, s, timo;
1476 	long *ijoblist;
1477 	struct aiocb **ujoblist;
1478 
1479 	if (uap->nent > AIO_LISTIO_MAX)
1480 		return EINVAL;
1481 
1482 	timo = 0;
1483 	if (uap->timeout) {
1484 		/* Get timespec struct. */
1485 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1486 			return error;
1487 
1488 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1489 			return (EINVAL);
1490 
1491 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1492 		if (itimerfix(&atv))
1493 			return (EINVAL);
1494 		timo = tvtohz_high(&atv);
1495 	}
1496 
1497 	ki = p->p_aioinfo;
1498 	if (ki == NULL)
1499 		return EAGAIN;
1500 
1501 	njoblist = 0;
1502 	ijoblist = zalloc(aiol_zone);
1503 	ujoblist = zalloc(aiol_zone);
1504 	cbptr = uap->aiocbp;
1505 
1506 	for (i = 0; i < uap->nent; i++) {
1507 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1508 		if (cbp == 0)
1509 			continue;
1510 		ujoblist[njoblist] = cbp;
1511 		ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1512 		njoblist++;
1513 	}
1514 
1515 	if (njoblist == 0) {
1516 		zfree(aiol_zone, ijoblist);
1517 		zfree(aiol_zone, ujoblist);
1518 		return 0;
1519 	}
1520 
1521 	error = 0;
1522 	for (;;) {
1523 		TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1524 			for (i = 0; i < njoblist; i++) {
1525 				if (((intptr_t)
1526 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1527 				    ijoblist[i]) {
1528 					if (ujoblist[i] != cb->uuaiocb)
1529 						error = EINVAL;
1530 					zfree(aiol_zone, ijoblist);
1531 					zfree(aiol_zone, ujoblist);
1532 					return error;
1533 				}
1534 			}
1535 		}
1536 
1537 		crit_enter();
1538 		for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1539 		    TAILQ_NEXT(cb, plist)) {
1540 			for (i = 0; i < njoblist; i++) {
1541 				if (((intptr_t)
1542 				    cb->uaiocb._aiocb_private.kernelinfo) ==
1543 				    ijoblist[i]) {
1544 					crit_exit();
1545 					if (ujoblist[i] != cb->uuaiocb)
1546 						error = EINVAL;
1547 					zfree(aiol_zone, ijoblist);
1548 					zfree(aiol_zone, ujoblist);
1549 					return error;
1550 				}
1551 			}
1552 		}
1553 
1554 		ki->kaio_flags |= KAIO_WAKEUP;
1555 		error = tsleep(p, PCATCH, "aiospn", timo);
1556 		crit_exit();
1557 
1558 		if (error == ERESTART || error == EINTR) {
1559 			zfree(aiol_zone, ijoblist);
1560 			zfree(aiol_zone, ujoblist);
1561 			return EINTR;
1562 		} else if (error == EWOULDBLOCK) {
1563 			zfree(aiol_zone, ijoblist);
1564 			zfree(aiol_zone, ujoblist);
1565 			return EAGAIN;
1566 		}
1567 	}
1568 
1569 /* NOTREACHED */
1570 	return EINVAL;
1571 #endif /* VFS_AIO */
1572 }
1573 
1574 /*
1575  * aio_cancel cancels any non-physio aio operations not currently in
1576  * progress.
1577  */
1578 int
1579 aio_cancel(struct aio_cancel_args *uap)
1580 {
1581 #ifndef VFS_AIO
1582 	return ENOSYS;
1583 #else
1584 	struct proc *p = curproc;
1585 	struct kaioinfo *ki;
1586 	struct aiocblist *cbe, *cbn;
1587 	struct file *fp;
1588 	struct filedesc *fdp;
1589 	struct socket *so;
1590 	struct proc *po;
1591 	int error;
1592 	int cancelled=0;
1593 	int notcancelled=0;
1594 	struct vnode *vp;
1595 
1596 	fdp = p->p_fd;
1597 	if ((u_int)uap->fd >= fdp->fd_nfiles ||
1598 	    (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1599 		return (EBADF);
1600 
1601         if (fp->f_type == DTYPE_VNODE) {
1602 		vp = (struct vnode *)fp->f_data;
1603 
1604 		if (vn_isdisk(vp,&error)) {
1605 			uap->sysmsg_result = AIO_NOTCANCELED;
1606         	        return 0;
1607 		}
1608 	} else if (fp->f_type == DTYPE_SOCKET) {
1609 		so = (struct socket *)fp->f_data;
1610 
1611 		crit_enter();
1612 
1613 		for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1614 			cbn = TAILQ_NEXT(cbe, list);
1615 			if ((uap->aiocbp == NULL) ||
1616 				(uap->aiocbp == cbe->uuaiocb) ) {
1617 				po = cbe->userproc;
1618 				ki = po->p_aioinfo;
1619 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1620 				TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1621 				TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1622 				if (ki->kaio_flags & KAIO_WAKEUP) {
1623 					wakeup(po);
1624 				}
1625 				cbe->jobstate = JOBST_JOBFINISHED;
1626 				cbe->uaiocb._aiocb_private.status=-1;
1627 				cbe->uaiocb._aiocb_private.error=ECANCELED;
1628 				cancelled++;
1629 /* XXX cancelled, knote? */
1630 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1631 				    SIGEV_SIGNAL)
1632 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1633 				if (uap->aiocbp)
1634 					break;
1635 			}
1636 		}
1637 		crit_exit();
1638 
1639 		if ((cancelled) && (uap->aiocbp)) {
1640 			uap->sysmsg_result = AIO_CANCELED;
1641 			return 0;
1642 		}
1643 	}
1644 	ki=p->p_aioinfo;
1645 	if (ki == NULL)
1646 		goto done;
1647 	crit_enter();
1648 
1649 	for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1650 		cbn = TAILQ_NEXT(cbe, plist);
1651 
1652 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1653 		    ((uap->aiocbp == NULL ) ||
1654 		     (uap->aiocbp == cbe->uuaiocb))) {
1655 
1656 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1657 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1658                                 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1659                                 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1660                                     plist);
1661 				cancelled++;
1662 				ki->kaio_queue_finished_count++;
1663 				cbe->jobstate = JOBST_JOBFINISHED;
1664 				cbe->uaiocb._aiocb_private.status = -1;
1665 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1666 /* XXX cancelled, knote? */
1667 			        if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1668 				    SIGEV_SIGNAL)
1669 					psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1670 			} else {
1671 				notcancelled++;
1672 			}
1673 		}
1674 	}
1675 	crit_exit();
1676 done:
1677 	if (notcancelled) {
1678 		uap->sysmsg_result = AIO_NOTCANCELED;
1679 		return 0;
1680 	}
1681 	if (cancelled) {
1682 		uap->sysmsg_result = AIO_CANCELED;
1683 		return 0;
1684 	}
1685 	uap->sysmsg_result = AIO_ALLDONE;
1686 
1687 	return 0;
1688 #endif /* VFS_AIO */
1689 }
1690 
1691 /*
1692  * aio_error is implemented in the kernel level for compatibility purposes only.
1693  * For a user mode async implementation, it would be best to do it in a userland
1694  * subroutine.
1695  */
1696 int
1697 aio_error(struct aio_error_args *uap)
1698 {
1699 #ifndef VFS_AIO
1700 	return ENOSYS;
1701 #else
1702 	struct proc *p = curproc;
1703 	struct aiocblist *cb;
1704 	struct kaioinfo *ki;
1705 	long jobref;
1706 
1707 	ki = p->p_aioinfo;
1708 	if (ki == NULL)
1709 		return EINVAL;
1710 
1711 	jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1712 	if ((jobref == -1) || (jobref == 0))
1713 		return EINVAL;
1714 
1715 	TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1716 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1717 		    jobref) {
1718 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1719 			return 0;
1720 		}
1721 	}
1722 
1723 	crit_enter();
1724 
1725 	for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1726 	    plist)) {
1727 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1728 		    jobref) {
1729 			uap->sysmsg_result = EINPROGRESS;
1730 			crit_exit();
1731 			return 0;
1732 		}
1733 	}
1734 
1735 	for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1736 	    plist)) {
1737 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1738 		    jobref) {
1739 			uap->sysmsg_result = EINPROGRESS;
1740 			crit_exit();
1741 			return 0;
1742 		}
1743 	}
1744 	crit_exit();
1745 
1746 	crit_enter();
1747 	for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1748 	    plist)) {
1749 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1750 		    jobref) {
1751 			uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1752 			crit_exit();
1753 			return 0;
1754 		}
1755 	}
1756 
1757 	for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1758 	    plist)) {
1759 		if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1760 		    jobref) {
1761 			uap->sysmsg_result = EINPROGRESS;
1762 			crit_exit();
1763 			return 0;
1764 		}
1765 	}
1766 	crit_exit();
1767 
1768 #if (0)
1769 	/*
1770 	 * Hack for lio.
1771 	 */
1772 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1773 	if (status == -1)
1774 		return fuword(&uap->aiocbp->_aiocb_private.error);
1775 #endif
1776 	return EINVAL;
1777 #endif /* VFS_AIO */
1778 }
1779 
1780 /* syscall - asynchronous read from a file (REALTIME) */
1781 int
1782 aio_read(struct aio_read_args *uap)
1783 {
1784 #ifndef VFS_AIO
1785 	return ENOSYS;
1786 #else
1787 	return aio_aqueue(uap->aiocbp, LIO_READ);
1788 #endif /* VFS_AIO */
1789 }
1790 
1791 /* syscall - asynchronous write to a file (REALTIME) */
1792 int
1793 aio_write(struct aio_write_args *uap)
1794 {
1795 #ifndef VFS_AIO
1796 	return ENOSYS;
1797 #else
1798 	return aio_aqueue(uap->aiocbp, LIO_WRITE);
1799 #endif /* VFS_AIO */
1800 }
1801 
1802 /* syscall - XXX undocumented */
1803 int
1804 lio_listio(struct lio_listio_args *uap)
1805 {
1806 #ifndef VFS_AIO
1807 	return ENOSYS;
1808 #else
1809 	struct proc *p = curproc;
1810 	int nent, nentqueued;
1811 	struct aiocb *iocb, * const *cbptr;
1812 	struct aiocblist *cb;
1813 	struct kaioinfo *ki;
1814 	struct aio_liojob *lj;
1815 	int error, runningcode;
1816 	int nerror;
1817 	int i;
1818 
1819 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1820 		return EINVAL;
1821 
1822 	nent = uap->nent;
1823 	if (nent > AIO_LISTIO_MAX)
1824 		return EINVAL;
1825 
1826 	if (p->p_aioinfo == NULL)
1827 		aio_init_aioinfo(p);
1828 
1829 	if ((nent + num_queue_count) > max_queue_count)
1830 		return EAGAIN;
1831 
1832 	ki = p->p_aioinfo;
1833 	if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1834 		return EAGAIN;
1835 
1836 	lj = zalloc(aiolio_zone);
1837 	if (!lj)
1838 		return EAGAIN;
1839 
1840 	lj->lioj_flags = 0;
1841 	lj->lioj_buffer_count = 0;
1842 	lj->lioj_buffer_finished_count = 0;
1843 	lj->lioj_queue_count = 0;
1844 	lj->lioj_queue_finished_count = 0;
1845 	lj->lioj_ki = ki;
1846 
1847 	/*
1848 	 * Setup signal.
1849 	 */
1850 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1851 		error = copyin(uap->sig, &lj->lioj_signal,
1852 		    sizeof(lj->lioj_signal));
1853 		if (error) {
1854 			zfree(aiolio_zone, lj);
1855 			return error;
1856 		}
1857 		if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1858 			zfree(aiolio_zone, lj);
1859 			return EINVAL;
1860 		}
1861 		lj->lioj_flags |= LIOJ_SIGNAL;
1862 		lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1863 	} else
1864 		lj->lioj_flags &= ~LIOJ_SIGNAL;
1865 
1866 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1867 	/*
1868 	 * Get pointers to the list of I/O requests.
1869 	 */
1870 	nerror = 0;
1871 	nentqueued = 0;
1872 	cbptr = uap->acb_list;
1873 	for (i = 0; i < uap->nent; i++) {
1874 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1875 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1876 			error = _aio_aqueue(iocb, lj, 0);
1877 			if (error == 0)
1878 				nentqueued++;
1879 			else
1880 				nerror++;
1881 		}
1882 	}
1883 
1884 	/*
1885 	 * If we haven't queued any, then just return error.
1886 	 */
1887 	if (nentqueued == 0)
1888 		return 0;
1889 
1890 	/*
1891 	 * Calculate the appropriate error return.
1892 	 */
1893 	runningcode = 0;
1894 	if (nerror)
1895 		runningcode = EIO;
1896 
1897 	if (uap->mode == LIO_WAIT) {
1898 		int command, found, jobref;
1899 
1900 		for (;;) {
1901 			found = 0;
1902 			for (i = 0; i < uap->nent; i++) {
1903 				/*
1904 				 * Fetch address of the control buf pointer in
1905 				 * user space.
1906 				 */
1907 				iocb = (struct aiocb *)
1908 				    (intptr_t)fuword(&cbptr[i]);
1909 				if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1910 				    == 0))
1911 					continue;
1912 
1913 				/*
1914 				 * Fetch the associated command from user space.
1915 				 */
1916 				command = fuword(&iocb->aio_lio_opcode);
1917 				if (command == LIO_NOP) {
1918 					found++;
1919 					continue;
1920 				}
1921 
1922 				jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1923 
1924 				TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1925 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1926 					    == jobref) {
1927 						if (cb->uaiocb.aio_lio_opcode
1928 						    == LIO_WRITE) {
1929 							p->p_stats->p_ru.ru_oublock
1930 							    +=
1931 							    cb->outputcharge;
1932 							cb->outputcharge = 0;
1933 						} else if (cb->uaiocb.aio_lio_opcode
1934 						    == LIO_READ) {
1935 							p->p_stats->p_ru.ru_inblock
1936 							    += cb->inputcharge;
1937 							cb->inputcharge = 0;
1938 						}
1939 						found++;
1940 						break;
1941 					}
1942 				}
1943 
1944 				crit_enter();
1945 				TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1946 					if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1947 					    == jobref) {
1948 						found++;
1949 						break;
1950 					}
1951 				}
1952 				crit_exit();
1953 			}
1954 
1955 			/*
1956 			 * If all I/Os have been disposed of, then we can
1957 			 * return.
1958 			 */
1959 			if (found == nentqueued)
1960 				return runningcode;
1961 
1962 			ki->kaio_flags |= KAIO_WAKEUP;
1963 			error = tsleep(p, PCATCH, "aiospn", 0);
1964 
1965 			if (error == EINTR)
1966 				return EINTR;
1967 			else if (error == EWOULDBLOCK)
1968 				return EAGAIN;
1969 		}
1970 	}
1971 
1972 	return runningcode;
1973 #endif /* VFS_AIO */
1974 }
1975 
1976 #ifdef VFS_AIO
1977 /*
1978  * This is a weird hack so that we can post a signal.  It is safe to do so from
1979  * a timeout routine, but *not* from an interrupt routine.
1980  */
1981 static void
1982 process_signal(void *aioj)
1983 {
1984 	struct aiocblist *aiocbe = aioj;
1985 	struct aio_liojob *lj = aiocbe->lio;
1986 	struct aiocb *cb = &aiocbe->uaiocb;
1987 
1988 	if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1989 	    (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1990 		psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1991 		lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1992 	}
1993 
1994 	if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1995 		psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
1996 }
1997 
1998 /*
1999  * Interrupt handler for physio, performs the necessary process wakeups, and
2000  * signals.
2001  */
2002 static void
2003 aio_physwakeup(struct buf *bp)
2004 {
2005 	struct aiocblist *aiocbe;
2006 	struct proc *p;
2007 	struct kaioinfo *ki;
2008 	struct aio_liojob *lj;
2009 
2010 	wakeup(bp);
2011 
2012 	aiocbe = (struct aiocblist *)bp->b_spc;
2013 	if (aiocbe) {
2014 		p = bp->b_caller1;
2015 
2016 		aiocbe->jobstate = JOBST_JOBBFINISHED;
2017 		aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2018 		aiocbe->uaiocb._aiocb_private.error = 0;
2019 		aiocbe->jobflags |= AIOCBLIST_DONE;
2020 
2021 		if (bp->b_flags & B_ERROR)
2022 			aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2023 
2024 		lj = aiocbe->lio;
2025 		if (lj) {
2026 			lj->lioj_buffer_finished_count++;
2027 
2028 			/*
2029 			 * wakeup/signal if all of the interrupt jobs are done.
2030 			 */
2031 			if (lj->lioj_buffer_finished_count ==
2032 			    lj->lioj_buffer_count) {
2033 				/*
2034 				 * Post a signal if it is called for.
2035 				 */
2036 				if ((lj->lioj_flags &
2037 				    (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2038 				    LIOJ_SIGNAL) {
2039 					lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2040 					callout_reset(&aiocbe->timeout, 0,
2041 							process_signal, aiocbe);
2042 				}
2043 			}
2044 		}
2045 
2046 		ki = p->p_aioinfo;
2047 		if (ki) {
2048 			ki->kaio_buffer_finished_count++;
2049 			TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2050 			TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2051 			TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2052 
2053 			KNOTE(&aiocbe->klist, 0);
2054 			/* Do the wakeup. */
2055 			if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2056 				ki->kaio_flags &= ~KAIO_WAKEUP;
2057 				wakeup(p);
2058 			}
2059 		}
2060 
2061 		if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2062 			callout_reset(&aiocbe->timeout, 0,
2063 					process_signal, aiocbe);
2064 		}
2065 	}
2066 }
2067 #endif /* VFS_AIO */
2068 
2069 /* syscall - wait for the next completion of an aio request */
2070 int
2071 aio_waitcomplete(struct aio_waitcomplete_args *uap)
2072 {
2073 #ifndef VFS_AIO
2074 	return ENOSYS;
2075 #else
2076 	struct proc *p = curproc;
2077 	struct timeval atv;
2078 	struct timespec ts;
2079 	struct kaioinfo *ki;
2080 	struct aiocblist *cb = NULL;
2081 	int error, s, timo;
2082 
2083 	suword(uap->aiocbp, (int)NULL);
2084 
2085 	timo = 0;
2086 	if (uap->timeout) {
2087 		/* Get timespec struct. */
2088 		error = copyin(uap->timeout, &ts, sizeof(ts));
2089 		if (error)
2090 			return error;
2091 
2092 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2093 			return (EINVAL);
2094 
2095 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2096 		if (itimerfix(&atv))
2097 			return (EINVAL);
2098 		timo = tvtohz_high(&atv);
2099 	}
2100 
2101 	ki = p->p_aioinfo;
2102 	if (ki == NULL)
2103 		return EAGAIN;
2104 
2105 	for (;;) {
2106 		if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2107 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2108 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2109 			if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2110 				p->p_stats->p_ru.ru_oublock +=
2111 				    cb->outputcharge;
2112 				cb->outputcharge = 0;
2113 			} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2114 				p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2115 				cb->inputcharge = 0;
2116 			}
2117 			aio_free_entry(cb);
2118 			return cb->uaiocb._aiocb_private.error;
2119 		}
2120 
2121 		crit_enter();
2122  		if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2123 			crit_exit();
2124 			suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2125 			uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2126 			aio_free_entry(cb);
2127 			return cb->uaiocb._aiocb_private.error;
2128 		}
2129 
2130 		ki->kaio_flags |= KAIO_WAKEUP;
2131 		error = tsleep(p, PCATCH, "aiowc", timo);
2132 		crit_exit();
2133 
2134 		if (error == ERESTART)
2135 			return EINTR;
2136 		else if (error < 0)
2137 			return error;
2138 		else if (error == EINTR)
2139 			return EINTR;
2140 		else if (error == EWOULDBLOCK)
2141 			return EAGAIN;
2142 	}
2143 #endif /* VFS_AIO */
2144 }
2145 
2146 #ifndef VFS_AIO
2147 static int
2148 filt_aioattach(struct knote *kn)
2149 {
2150 
2151 	return (ENXIO);
2152 }
2153 
2154 struct filterops aio_filtops =
2155 	{ 0, filt_aioattach, NULL, NULL };
2156 
2157 #else
2158 /* kqueue attach function */
2159 static int
2160 filt_aioattach(struct knote *kn)
2161 {
2162 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2163 
2164 	/*
2165 	 * The aiocbe pointer must be validated before using it, so
2166 	 * registration is restricted to the kernel; the user cannot
2167 	 * set EV_FLAG1.
2168 	 */
2169 	if ((kn->kn_flags & EV_FLAG1) == 0)
2170 		return (EPERM);
2171 	kn->kn_flags &= ~EV_FLAG1;
2172 
2173 	SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2174 
2175 	return (0);
2176 }
2177 
2178 /* kqueue detach function */
2179 static void
2180 filt_aiodetach(struct knote *kn)
2181 {
2182 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2183 
2184 	SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2185 }
2186 
2187 /* kqueue filter function */
2188 /*ARGSUSED*/
2189 static int
2190 filt_aio(struct knote *kn, long hint)
2191 {
2192 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2193 
2194 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2195 	if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2196 	    aiocbe->jobstate != JOBST_JOBBFINISHED)
2197 		return (0);
2198 	kn->kn_flags |= EV_EOF;
2199 	return (1);
2200 }
2201 
2202 struct filterops aio_filtops =
2203 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
2204 #endif /* VFS_AIO */
2205